Skip to content

Getting Started

Use this section to set up the runtime environment and launch the first asynchronous training workflow.

  • Installation

    Prepare the verl-compatible Python environment used by Claw-R1.

    Installation

  • Quick Start

    Run a black-box GSM8K training example and inspect it with the dashboard.

    Quick Start

Requirements

Dependency Recommended baseline
Python 3.10+
PyTorch 2.0+
CUDA 12.1+
Ray 2.10+
GPU At least 3 GPUs for the small async example

Runtime Shape

Agent -> Gateway -> DataPool -> Trainer
                     ^             |
                     |             v
                  Dashboard <- Parameter Synchronizer

The Gateway receives agent traffic, DataPool stores step-level data, the Trainer consumes curated batches, and the dashboard monitors the live data lifecycle.